3D U-Net assisted fault probability prediction from seismic volume in Dibrugarh oil field, upper Assam shelf, NE India

We presented a novel 3D U-Net Deep Convolutional Neural Network learning-based workflow for predicting the fault probability network from a 3D seismic volume associated with the geologically complex petroliferous basin. The workflow begins with applying a Dip-steered Median Filter (DSMF) to clean the seismic data, followed by Edge-Preserving Smoother (EPS) filter to enhance fault localisation. Subsequently, the Fault Likelihood (FL) attribute is computed from the EPS-filtered volume, followed by the Thin Fault Likelihood (TFL) attribute computation from the FL volume. Further, a fault mask volume was generated through a conditional mathematical attribute from the TFL volume. Two separate deep learning models were trained to predict fault probability networks. The first one utilised DSMF-filtered seismic volumes as input, and the second used EPS-enhanced seismic volumes, while in both cases the corresponding fault mask volume was set as the target. The feasibility of the proposed workflow was tested using 3D seismic data from the Dibrugarh field of the Upper Assam Shelf, India. Both models achieve >85 % accuracy in the training phase and accurately predict faults in the test phase. The EPS volume-based model has a higher accuracy of 89 %. The predicted fault volumes are then passed through a skeletonization filter for more accurate fault localisation. The demonstrated novel fault probability prediction process can predict faults from the voluminous seismic data from structurally complex geological settings with higher accuracy and less computational time. It can apply to the E&P industry's automated subsurface structural interpretation.